With the rapid development of power Internet of Things (IoT) scenarios such as smart factories and smart homes, numerous intelligent terminal devices and real-time interactive applications impose higher demands on computing latency and resource supply efficiency. Multi-access edge computing technology deploys cloud computing capabilities at the network edge; constructs distributed computing nodes and multi-access systems and offers infrastructure support for services with low latency and high reliability. Existing research relies on a strong assumption that the environmental state is fully observable and fails to thoroughly consider the continuous time-varying features of edge server load fluctuations, leading to insufficient adaptability of the model in a heterogeneous dynamic environment. Thus, this paper establishes a framework for end-edge collaborative task offloading based on a partially observable Markov decision-making process (POMDP) and proposes a method for end-edge collaborative task offloading in heterogeneous scenarios. It achieves time-series modeling of the historical load characteristics of edge servers and endows the agent with the ability to be aware of the load in dynamic environmental states. Moreover, by dynamically assessing the exploration value of historical trajectories in the central trajectory pool and adjusting the sample weight distribution, directional exploration and strategy optimization of high-value trajectories are realized. Experimental results indicate that the proposed method exhibits distinct advantages compared with existing methods in terms of average delay and task failure rate and also verifies the method’s robustness in a dynamic environment.

With the deep integration of edge computing, 5G and Artificial Intelligence of Things (AIoT) technologies, the large-scale deployment of intelligent terminal devices has given rise to data silos and privacy security challenges in sensing-computing fusion scenarios. Traditional federated learning (FL) algorithms face significant limitations in practical applications due to client drift, model bias, and resource constraints under non-independent and identically distributed (Non-IID) data, as well as the computational overhead and utility loss caused by privacy-preserving techniques. To address these issues, this paper proposes an Efficient and Privacy-enhancing Clustering Federated Learning method (FedEPC). This method introduces a dual-round client selection mechanism to optimize training. First, the Sparsity-based Privacy-preserving Representation Extraction Module (SPRE) and Adaptive Isomorphic Devices Clustering Module (AIDC) cluster clients based on privacy-sensitive features. Second, the Context-aware In-cluster Client Selection Module (CICS) dynamically selects representative devices for training, ensuring heterogeneous data distributions are fully represented. By conducting federated training within clusters and aggregating personalized models, FedEPC effectively mitigates weight divergence caused by data heterogeneity, reduces the impact of client drift and straggler issues. Experimental results demonstrate that FedEPC significantly improves test accuracy in highly Non-IID data scenarios compared to FedAvg and existing clustering FL methods. By ensuring privacy security, FedEPC provides an efficient and robust solution for FL in resource-constrained devices within sensing-computing fusion scenarios, offering both theoretical value and engineering practicality.